A Holistically-Guided Decoder for Deep Representation Learning with Applications to Semantic Segmentation and Object Detection.